README.md

---
license: apache-2.0
tags:
- devsecops
- llm
- sft
- lora
- tulu-3
- kubernetes
- terraform
---

# DevSecOps Model Platform

> Train a secure model on the best data, then deploy it securely.

## Start Here: Train Your Model

| Dataset | Size | What It Gives You | Command |
|---------|------|-------------------|---------|
| **tulu-3-sft-mixture** | 940K | Math, code, safety, chat (BEST) | python model/train_tulu3.py |
| **OpenThoughts-114k** | 114K | Reasoning, chain-of-thought | python model/train_openthoughts.py |

**allenai/tulu-3-sft-mixture** is from Allen AI Tulu 3 - current SOTA open instruction-tuned model. Proven on Llama-3.1-8B: MMLU 53.5, GSM8K 79.9, HumanEval 76.8.

LoRA config from LoRA Without Regret (Schulman 2025): r=256, alpha=16, all-linear = matches full fine-tuning at 67% compute.

## Repository Structure

```
model/                     THE MODEL - train, serve, enhance
  train_tulu3.py             Primary: 940K best data (zero preprocessing)
  train_openthoughts.py      Reasoning: 114K CoT traces
  finetune_configurable.py   Multi-dataset configurable trainer
  rag_pipeline.py             RAG for DevSecOps knowledge
  DATASETS.md                 Why these datasets, proven recipes

deployment/               SERVE IT - Kubernetes + Docker + vLLM
  deployment.yaml             ML inference K8s manifest
  mlflow-deployment.yaml      Experiment tracking
  Dockerfile.ml-inference     Hardened multi-stage image

security/                 PROTECT IT - scanning + policies
  scanning/                   Trivy, Semgrep, Checkov, SBOM
  policies/                   Kyverno, OPA Gatekeeper

infrastructure/           RUN IT - Terraform + monitoring + CI/CD
  terraform/                  VPC, EKS, RDS, S3, IAM, KMS, GuardDuty, Macie
  monitoring/                 Prometheus, Alertmanager, OTEL, Grafana
  ci-cd/                      GitHub Actions DevSecOps pipeline

compliance/               CERTIFY IT - SOC2, NIST, CIS
  controls-mapping.yaml       SOC2 Type II
  nist-800-53-mapping.yaml    NIST 800-53 Rev5
  cis-eks-k8s.yaml            CIS Benchmarks
```

## Quick Commands

```bash
# Train on best data (A100, ~6h)
python model/train_tulu3.py

# Quick test (any GPU)
python model/train_tulu3.py --max_steps 100 --no_push

# Security scan
python security/scanning/security_audit.py

# Deploy model to K8s
kubectl apply -f deployment/deployment.yaml

# Infrastructure (Terraform)
cd infrastructure/terraform/environments/prod && terraform apply
```